Update of /cvsroot/pure-data/externals/pdp/system/mmx In directory sc8-pr-cvs1.sourceforge.net:/tmp/cvs-serv6756/system/mmx
Added Files: Makefile pdp_mmx_test.c pixel_add_s16.s pixel_biquad_dirI_s16.s pixel_biquad_s16.s pixel_ca_s1.s pixel_cascade_s16.s pixel_cheby_s16.s pixel_conv_hor_s16.s pixel_conv_ver_s16.s pixel_crot_s16.s pixel_gain.s pixel_gain_s16.s pixel_mix_s16.s pixel_mul_s16.s pixel_pack_s16u8.s pixel_rand_s16.s pixel_randmix_s16.s pixel_resample_s16.s pixel_s1.s pixel_unpack_u8s16.s Log Message: checking in pdp 0.12.4 from http://zwizwa.fartit.com/pd/pdp/pdp-0.12.4.tar.gz
--- NEW FILE: pixel_resample_s16.s ---
#interpolation data: #* 4 vectors: neighbourhood for samples (TL, TR, BL, BR) #* 2 vectors: fractional part (unsigned) #* 2 vectors: addresses of pixel blocks
#coord conversion data: #1 vector: 32bit splatted address #1 vector: 16bit splatted w-1 #1 vector: 16bit splatted h-1 #1 vector: 16bit splatted w (reuse w-1 with add?) #1 dword: 32 bit line offset
#coord generation data: several vectors for parameter update stuff..
#coordinate systems: 16 bit virtual coordinates (signed, center relative) #* 2 vectors: virtual coordinates #(evt tussenstap + conversie naar 16 bit virtual)
#step 1: generate virtual coords
#step 2: virtual coords -> block adresses + fractional adresses #* mulhigh: real coords (x,y) (center relative) #* add center -> unsigned (top left relative) #* mullow: fractional part (x_frac, y_frac) #* mulhigh, mullow, pack 32bit: y_offset #* pack 32bit: x_offset #* add, shift, add start address: real addresses
#step3: data fetch using generated addresses: # this step would be much simpler in 4x16bit rgba. life's a bitch..
#step4: billinear interpolation
#stat5: store
# this can be simplified by doing 32 bit unaligned moves # and vector unpacking on the data
# cooked image data structure # pixel environment temp storage TL1 = 0x00 TL2 = 0x02 TL3 = 0x04 TL4 = 0x06 TR1 = 0x08 TR2 = 0x0A TR3 = 0x0C TR4 = 0x0E BL1 = 0x10 BL2 = 0x12 BL3 = 0x14 BL4 = 0x16 BR1 = 0x18 BR2 = 0x1A BR3 = 0x1C BR4 = 0x1E # addresses of pixel blocks ADDRESS1 = 0x20 ADDRESS2 = 0x24 ADDRESS3 = 0x28 ADDRESS4 = 0x2C
# second env + address buffer (testing: not used) SECONDBUFFER = 0x30 # 32bit splatted bitmap address V2PLANEADDRESS = 0x60 # 16bit splatted image constants V4TWOWIDTHM1 = 0x68 V4TWOHEIGHTM1 = 0x70 V4LINEOFFSET = 0x78 # data struct size RESAMPLEDATASIZE = 0x80
# interpolation routine # input: %mm0, %mm1 4 x 16bit unsigned top left relative virtual x and y coordinates # %esi: temp & algo data structure
getpixelsbilin: psrlw $1, %mm0 # convert to range 0->0x7fff [0,0.5[ psrlw $1, %mm1 movq %mm0, %mm2 movq %mm1, %mm3 movq V4TWOWIDTHM1(%esi), %mm4 # 2 * (width - 1) movq V4TWOHEIGHTM1(%esi), %mm5 # 2 * (height - 1) pmulhw %mm5, %mm3 # mm3 == y coord (topleft relative) pmulhw %mm4, %mm2 # mm2 == x coord (topleft relative) pmullw %mm5, %mm1 # mm1 == y frac (unsigned) pmullw %mm4, %mm0 # mm0 == x frac (unsigned)
movq %mm3, %mm5 # copy y coord pmullw V4LINEOFFSET(%esi), %mm3 # low part of line offset pmulhw V4LINEOFFSET(%esi), %mm5 # high part of line offset
movq %mm2, %mm7 # copy x coord vector pxor %mm4, %mm4 punpcklwd %mm4, %mm2 # low part in %mm2 punpckhwd %mm4, %mm7 # hight part in %mm7 movq %mm3, %mm6 # copy punpcklwd %mm5, %mm3 # unpack low part in %mm3 punpckhwd %mm5, %mm6 # high part int %mm6
paddd %mm2, %mm3 paddd %mm7, %mm6 pslld $1, %mm3 # convert to word adresses pslld $1, %mm6
paddd V2PLANEADDRESS(%esi), %mm3 # add pixel plane address paddd V2PLANEADDRESS(%esi), %mm6
movq %mm3, ADDRESS1(%esi) # store adresses movq %mm6, ADDRESS3(%esi)
pcmpeqw %mm2, %mm2 # all ones movq %mm0, %mm4 # copy x frac movq %mm1, %mm5 # copy y frac pxor %mm2, %mm4 # compute compliment (approx negative) pxor %mm2, %mm5
psrlw $1, %mm0 # shift right (0.5 * (frac x) psrlw $1, %mm1 # shift right (0.5 * (frac y) psrlw $1, %mm4 # shift right (0.5 * (1 - frac x) psrlw $1, %mm5 # shift right (0.5 * (1 - frac y)
movq %mm0, %mm2 # copy of frac x movq %mm4, %mm3 # copy of (1-frac x) # fetch data
#jmp skipfetch # seems the fetch is the real killer. try to optimize this # using 32 bit accesses & shifts
# the src image data struct is padded to the cooked data struct movl RESAMPLEDATASIZE(%esi), %edi shll $1, %edi
movl ADDRESS1(%esi), %ecx movl ADDRESS2(%esi), %edx movw (%ecx), %ax movw (%edx), %bx movw %ax, TL1(%esi) movw %bx, TL2(%esi) movw 2(%ecx), %ax movw 2(%edx), %bx movw %ax, TR1(%esi) movw %bx, TR2(%esi)
addl %edi, %ecx addl %edi, %edx
movw (%ecx), %ax movw (%edx), %bx movw %ax, BL1(%esi) movw %bx, BL2(%esi) movw 2(%ecx), %ax movw 2(%edx), %bx movw %ax, BR1(%esi) movw %bx, BR2(%esi)
movl ADDRESS3(%esi), %ecx movl ADDRESS4(%esi), %edx
movw (%ecx), %ax movw (%edx), %bx movw %ax, TL3(%esi) movw %bx, TL4(%esi) movw 2(%ecx), %ax movw 2(%edx), %bx movw %ax, TR3(%esi) movw %bx, TR4(%esi) addl %edi, %ecx addl %edi, %edx
movw (%ecx), %ax movw (%edx), %bx movw %ax, BL3(%esi) movw %bx, BL4(%esi) movw 2(%ecx), %ax movw 2(%edx), %bx movw %ax, BR3(%esi) movw %bx, BR4(%esi)
skipfetch: pmulhw TL1(%esi), %mm4 # bilin interpolation pmulhw TR1(%esi), %mm0 pmulhw BL1(%esi), %mm3 pmulhw BR1(%esi), %mm2
paddw %mm4, %mm0 paddw %mm3, %mm2
pmulhw %mm5, %mm0 pmulhw %mm1, %mm2
paddw %mm2, %mm0 psllw $2, %mm0 # compensate for gain reduction
ret
// linear mapping data struct ROWSTATEX = 0x0 ROWSTATEY = 0x8 COLSTATEX = 0x10 COLSTATEY = 0x18 ROWINCX = 0x20 ROWINCY = 0x28 COLINCX = 0x30 COLINCY = 0x38
// image data struct LINEOFFSET = 0x0 IMAGEADDRESS = 0x4 WIDTH = 0x8 HEIGHT = 0xC IMAGEDATASIZE = 0x10
# pixel_resample_linmap_s16(void *x) .globl pixel_resample_linmap_s16 .type pixel_resample_linmap_s16,@function
SOURCEIMAGE = RESAMPLEDATASIZE DESTIMAGE = SOURCEIMAGE + IMAGEDATASIZE LINMAPDATA = DESTIMAGE + IMAGEDATASIZE pixel_resample_linmap_s16: pushl %ebp movl %esp, %ebp pushl %esi pushl %edi pushl %ebx
movl 8(%ebp), %esi # get data struct movl DESTIMAGE+HEIGHT(%esi), %edx # image height movl DESTIMAGE+IMAGEADDRESS(%esi), %edi # dest image address movl DESTIMAGE+WIDTH(%esi), %ecx # image width shrl $2, %ecx # vector count .align 16 linmap_looprow: movq LINMAPDATA+ROWSTATEX(%esi), %mm0 # get current coordinates movq LINMAPDATA+ROWSTATEY(%esi), %mm1
linmap_loopcol: movq %mm0, %mm4 # copy movq %mm1, %mm5 paddd LINMAPDATA+ROWINCX(%esi), %mm4 # increment paddd LINMAPDATA+ROWINCY(%esi), %mm5 movq %mm4, %mm6 # copy movq %mm5, %mm7 paddd LINMAPDATA+ROWINCX(%esi), %mm6 # increment paddd LINMAPDATA+ROWINCY(%esi), %mm7 movq %mm6, LINMAPDATA+ROWSTATEX(%esi) # store next state movq %mm7, LINMAPDATA+ROWSTATEY(%esi)
psrad $16, %mm0 # round to 16 bit psrad $16, %mm1 psrad $16, %mm4 psrad $16, %mm5 packssdw %mm4, %mm0 # pack new coordinates packssdw %mm5, %mm1 push %ecx push %edx push %edi call getpixelsbilin # do interpolation
pop %edi pop %edx pop %ecx movq %mm0, (%edi) # store 4 pixels addl $0x8, %edi # point to next 4 pixels decl %ecx # dec row counter jnz linmap_looprow
movq LINMAPDATA+COLSTATEX(%esi), %mm0 # get column state vector movq LINMAPDATA+COLSTATEY(%esi), %mm1 movl DESTIMAGE+WIDTH(%esi), %ecx # image width shrl $2, %ecx # vector count paddd LINMAPDATA+COLINCX(%esi), %mm0 # increment paddd LINMAPDATA+COLINCY(%esi), %mm1 movq %mm0, LINMAPDATA+COLSTATEX(%esi) # store movq %mm1, LINMAPDATA+COLSTATEY(%esi) decl %edx # dec column counter jnz linmap_loopcol emms popl %ebx popl %edi popl %esi leave ret
--- NEW FILE: pixel_gain_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # .globl pixel_gain_s16 .type pixel_gain_s16,@function
# gain is integer, shift count is down # void pixel_gain_s16(int *buf, int nb_8pixel_vectors, short int gain[4], unsigned long long *shift)
pixel_gain_s16: pushl %ebp movl %esp, %ebp push %esi push %edi
movl 20(%ebp), %edi movq (%edi), %mm6 # get shift vector
movl 16(%ebp), %edi movq (%edi), %mm7 # get gain vector movl 8(%ebp), %esi # input array movl 12(%ebp), %ecx # pixel count
.align 16 .loop_gain:
movq (%esi), %mm0 # load 4 pixels from memory movq %mm0, %mm1 pmulhw %mm7, %mm1 # apply gain (s15.0) fixed point, high word pmullw %mm7, %mm0 # low word
movq %mm0, %mm2 # copy movq %mm1, %mm3
punpcklwd %mm1, %mm0 # unpack lsw components punpckhwd %mm3, %mm2 # unpack msw components
psrad %mm6, %mm0 # apply signed shift psrad %mm6, %mm2
packssdw %mm2, %mm0 # pack result & saturate movq %mm0, (%esi) # store result
addl $8, %esi # increment source pointer decl %ecx jnz .loop_gain # loop
emms pop %edi pop %esi leave ret
--- NEW FILE: pixel_cheby_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # .globl pixel_cheby_s16_3plus .type pixel_cheby_s16_3plus,@function
# void pixel_cheby_s16(int *buf, int nb_8pixel_vectors, int order+1, short int *coefs)
# coefs are s2.13 fixed point (-4->4) pixel_cheby_s16_3plus: pushl %ebp movl %esp, %ebp push %esi push %edi push %edx
movl 8(%ebp), %esi # input array movl 12(%ebp), %ecx # vector count movl 16(%ebp), %eax # get order+1
shll $3, %eax movl 20(%ebp), %edx addl %eax, %edx # edx = coef endx address # jmp skip .align 16 .loop_cheby:
movl 20(%ebp), %edi # get coefs movq (%esi), %mm0 # load 4 pixels from memory (mm0 = x) pcmpeqw %mm2, %mm2 movq %mm0, %mm1 # mm1 (T_n-1) <- x psrlw $1, %mm2 # mm2 (T_n-2) <- 1
movq (%edi), %mm4 # mm4 (acc) == a0 psraw $1, %mm4 # mm4 == a0/2 movq %mm0, %mm5 # mm5 (intermediate) pmulhw 8(%edi), %mm5 # mm5 == (x * a1)/2 paddsw %mm5, %mm4 # acc = c0 + c1 x addl $16, %edi
.loop_cheby_inner: movq %mm1, %mm3 # mm3 == T_n-1 psraw $2, %mm2 # mm2 == T_n-2 / 4 pmulhw %mm0, %mm3 # mm3 == (2 x T_n-1) / 4 psubsw %mm2, %mm3 # mm3 == (2 x T_n-1 - T_n-2) / 4 paddsw %mm3, %mm3 paddsw %mm3, %mm3 # mm3 == T_n movq %mm1, %mm2 # mm2 == new T_n-1 movq %mm3, %mm1 # mm3 == new T_n-2 pmulhw (%edi), %mm3 # mm3 = a_n * T_n / 2 paddsw %mm3, %mm4 # accumulate addl $8, %edi cmpl %edx, %edi jne .loop_cheby_inner paddsw %mm4, %mm4 # compensate for 0.125 factor paddsw %mm4, %mm4 paddsw %mm4, %mm4 movq %mm4, (%esi) # store result in memory addl $8, %esi # increment source/dest pointer decl %ecx jnz .loop_cheby # loop
skip: emms
pop %edx pop %edi pop %esi leave ret
--- NEW FILE: pixel_cascade_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. #
# TODO: COUPLED CASCADE SECOND ORDER SECTION # # s1[k] = ar * s1[k-1] + ai * s2[k-1] + x[k] # s2[k] = ar * s2[k-1] - ai * s1[k-1] # y[k] = c0 * x[k] + c1 * s1[k-1] + c2 * s2[k-1]
# MACRO: df2 # # computes a coupled cascade # # input: %mm0 == input # %mm1 == state 1 # %mm2 == state 2 # (%esi) == cascade coefs (ar ai c0 c1 c2) in s0.15 # output: %mm0 == output # %mm1 == state 1 # %mm2 == state 2
.macro coupled pmovq %mm1, %mm3 # mm3 == s1[k-1] pmovq %mm1, %mm4 # mm4 == s1[k-1] pmovq %mm2, %mm5 # mm5 == s2[k-1] pmovq %mm2, %mm6 # mm5 == s2[k-1] pmulhw (%esi), %mm1 # mm1 == s1[k-1] * ar pmulhw 8(%esi), %mm3 # mm3 == s1[k-1] * ai pmulhw 24(%esi), %mm4 # mm4 == s1[k-1] * c1 pmulhw (%esi), %mm2 # mm2 == s2[k-1] * ar pmulhw 8(%esi), %mm5 # mm5 == s2[k-1] * ai pmulhw 32(%esi), %mm6 # mm6 == s2[k-1] * c2 paddw %mm5, %mm1 # mm1 == s1[k-1] * ar + s2[k-1] * ai psubw %mm3, %mm2 # mm2 == s2[k-1] * ar - s1[k-1] * ai == s2[k] paddw %mm0, %mm1 # mm1 == s1[k] pmulhw 16(%esi), %mm0 # mm0 == x[k] * c0 paddw %mm6, %mm4 # mm4 == s1[k-1] * c1 + s2[k-1] * c2 paddw %mm4, %mm0 # mm0 == y[k] .endm
# in order to use the 4 line parallel cascade routine on horizontal # lines, we need to reorder (rotate or transpose) the matrix, since # images are scanline encoded, and we want to work in parallell # on 4 lines. # # since the 4 lines are independent, it doesnt matter in which order # the the vector elements are present. # # this allows us to use the same routine for left->right and right->left # processing. # # some comments on the non-abelean group of square isometries consisting of # (I) identity # (H) horizontal axis mirror # (V) vertical axis mirror # (T) transpose (diagonal axis mirror) # (A) antitranspose (antidiagonal axis mirror) # (R1) 90deg anticlockwize rotation # (R2) 180deg rotation # (R3) 90deg clockwize rotation # # # we basicly have two options: (R1,R3) or (T,A) # we opt for T and A because they are self inverting, which improves locality # # use antitranspose for right to left an transpose # for left to right (little endian)
# antitranspose 4x4
# input # %mm3 == {d0 d1 d2 d3} # %mm2 == {c0 c1 c2 c3} # %mm1 == {b0 b1 b2 b3} # %mm0 == {a0 a1 a2 a3}
# output # %mm3 == {a3 b3 c3 d3} # %mm2 == {a2 b2 c2 d2} # %mm1 == {a1 b1 c1 d1} # %mm0 == {a0 b0 c0 d0}
.macro antitranspose_4x4: movq %mm3, %mm4 punpcklwd %mm1, %mm4 # mm4 <- {b2 d2 b3 d3} movq %mm3, %mm5 punpckhwd %mm1, %mm5 # mm5 <- {b0 d0 b1 d1} movq %mm2, %mm6 punpcklwd %mm0, %mm6 # mm6 <- {a2 c2 a3 c3} movq %mm2, %mm7 punpckhwd %mm0, %mm7 # mm7 <- {a0 c0 a1 c1}
movq %mm4, %mm3 punpcklwd %mm6, %mm3 # mm3 <- {a3 b3 c3 d3} movq %mm4, %mm2 punpckhwd %mm6, %mm2 # mm2 <- {a2 b2 c2 d2} movq %mm5, %mm1 punpcklwd %mm7, %mm1 # mm1 <- {a1 b1 c1 d1} movq %mm5, %mm0 punpckhwd %mm7, %mm0 # mm0 <- {a0 b0 c0 d0} .endm
# transpose 4x4
# input # %mm3 == {d3 d2 d1 d0} # %mm2 == {c3 c2 c1 c0} # %mm1 == {b3 b2 b1 b0} # %mm0 == {a3 a2 a1 a0}
# output # %mm3 == {d3 c3 b3 a3} # %mm2 == {d2 c2 b2 a2} # %mm1 == {d1 c1 b1 a1} # %mm0 == {d0 c0 b0 a0}
.macro transpose_4x4: movq %mm0, %mm4 punpcklwd %mm2, %mm4 # mm4 <- {c1 a1 c0 a0} movq %mm0, %mm5 punpckhwd %mm2, %mm5 # mm5 <- {c3 a3 c2 a2} movq %mm1, %mm6 punpcklwd %mm3, %mm6 # mm6 <- {d1 b1 d0 b0} movq %mm1, %mm7 punpckhwd %mm3, %mm7 # mm7 <- {d3 b3 d2 b2}
movq %mm4, %mm0 punpcklwd %mm6, %mm0 # mm0 <- {d0 c0 b0 a0} movq %mm4, %mm1 punpckhwd %mm6, %mm1 # mm1 <- {d1 c1 b1 a1} movq %mm5, %mm2 punpcklwd %mm7, %mm2 # mm2 <- {d2 c2 b2 a2} movq %mm5, %mm3 punpckhwd %mm7, %mm3 # mm3 <- {d3 c3 b3 a3}
.endm .globl pixel_cascade_vertb_s16 .type pixel_cascade_vertb_s16,@function
# pixel_cascade_vertbr_s16(char *pixel_array, int nb_rows, int linewidth, short int coef[20], short int state[8])
pixel_cascade_vertb_s16:
pushl %ebp movl %esp, %ebp push %ebx push %esi push %edi
movl 8(%ebp), %ebx # pixel array offset movl 12(%ebp), %ecx # nb of 4x4 pixblocks movl 16(%ebp), %edx # line with
movl 20(%ebp), %esi # coefs movl 24(%ebp), %edi # state
shll $1, %edx # short int addressing subl %edx, %ebx
movq 0(%edi), %mm1 # s1[k-1] movq 8(%edi), %mm2 # s2[k-1] .align 16 .cascade_vertb_line_loop: movq (%ebx,%edx,1), %mm3 movq %mm3, %mm0 addl %edx, %ebx coupled movq %mm0, (%ebx) movq (%ebx,%edx,1), %mm3 movq %mm3, %mm0 addl %edx, %ebx coupled movq %mm0, (%ebx) movq (%ebx,%edx,1), %mm3 movq %mm3, %mm0 addl %edx, %ebx coupled movq %mm0, (%ebx) movq (%ebx,%edx,1), %mm3 movq %mm3, %mm0 addl %edx, %ebx coupled movq %mm0, (%ebx) decl %ecx jnz .cascade_vertb_line_loop movq %mm1, 0(%edi) # s1[k-1] movq %mm2, 8(%edi) # s2[k-1]
emms pop %edi pop %esi pop %ebx leave ret
.globl pixel_cascade_horlr_s16 .type pixel_cascade_horlr_s16,@function
# pixel_cascade_hor_s16(char *pixel_array, int nb_rows, int linewidth, short int coef[20], short int state[8])
pixel_cascade_horlr_s16:
pushl %ebp movl %esp, %ebp push %ebx push %esi push %edi
movl 8(%ebp), %ebx # pixel array offset movl 12(%ebp), %ecx # nb of 4x4 pixblocks movl 16(%ebp), %edx # line with
movl 20(%ebp), %esi # coefs movl 24(%ebp), %edi # state
shll $1, %edx # short int addressing movl %edx, %eax shll $1, %eax addl %edx, %eax # eax = 3 * edx
.align 16 .cascade_horlr_line_loop: movq (%edi), %mm1 movq 8(%edi), %mm2 movq (%ebx), %mm0 movq (%ebx,%edx,1), %mm1 movq (%ebx,%edx,2), %mm2 movq (%ebx,%eax,1), %mm3 transpose_4x4 movq %mm1, (%ebx,%edx,1) movq %mm2, (%ebx,%edx,2) movq %mm3, (%ebx,%eax,1)
coupled
movq %mm0, (%ebx) movq (%ebx,%edx,1), %mm3 movq %mm3, %mm0
coupled
movq %mm0, (%ebx, %edx,1) movq (%ebx,%edx,2), %mm3 movq %mm3, %mm0
coupled
movq %mm0, (%ebx, %edx,2) movq (%ebx,%eax,1), %mm3 movq %mm3, %mm0
coupled movq %mm1, 0(%edi) # s1[k-1] movq %mm2, 8(%edi) # s2[k-1]
movq %mm0, %mm3 movq (%ebx), %mm0 movq (%ebx,%edx,1), %mm1 movq (%ebx,%edx,2), %mm2
transpose_4x4 movq %mm0, (%ebx) movq %mm1, (%ebx,%edx,1) movq %mm2, (%ebx,%edx,2) movq %mm3, (%ebx,%eax,1)
addl $8, %ebx decl %ecx jnz .cascade_horlr_line_loop emms pop %edi pop %esi pop %ebx leave ret
--- NEW FILE: pixel_rand_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # .globl pixel_rand_s16 .type pixel_rand_s16,@function
# mmx rgba pixel gain # void pixel_rand_s16(int *dst, nb_4pixel_vectors, short int random_seed[4])
pixel_rand_s16: pushl %ebp movl %esp, %ebp push %esi push %edi
movl 16(%ebp), %esi # int16[4] array of random seeds movl 8(%ebp), %edi # dst array movl 12(%ebp), %ecx # pixel count
movq (%esi), %mm6
pcmpeqw %mm3, %mm3 psrlw $15, %mm3 # get bit mask 4 times 0x0001 .align 16 .loop_rand:
# prefetch 128(%esi)
movq %mm6, %mm4 # get random vector psrlw $15, %mm4 # get first component movq %mm6, %mm5 psrlw $14, %mm5 # get second component pxor %mm5, %mm4 movq %mm6, %mm5 psrlw $12, %mm5 # get third component pxor %mm5, %mm4 movq %mm6, %mm5 psrlw $3, %mm5 # get forth component pxor %mm5, %mm4
psllw $1, %mm6 # shift left original random vector pand %mm3, %mm4 # isolate new bit por %mm4, %mm6 # combine into new random vector
movq %mm6, (%edi) addl $8, %edi decl %ecx jnz .loop_rand # loop
movq %mm6, (%esi) # store random seeds
emms pop %edi pop %esi leave ret
--- NEW FILE: pixel_conv_ver_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # #TODO: fix out of bound acces in conv_ver and conv_hor # intermediate function # input in register: # %mm0: top 4 pixels # %mm1: middle 4 pixels # %mm2: bottom 4 pixels
# %mm5: top 4 pixel mask # %mm6: middle 4 pixel mask # %mm7: bottom 4 pixel mask # return in register: # %mm0: middle 4 pixels result
.conv_ver_4_pixels: .align 16 # compute quadruplet
# get top pixel pmulhw %mm5, %mm0 psllw $1, %mm0 # get middle pixel movq %mm1, %mm4 pmulhw %mm6, %mm4 psllw $1, %mm4 paddsw %mm4, %mm0
# get bottom pixel movq %mm2, %mm3 pmulhw %mm7, %mm3 psllw $1, %mm3 # mm3 <- mm3/4 paddsw %mm3, %mm0
ret .globl pixel_conv_ver_s16 .type pixel_conv_ver_s16,@function
# pixel_conv_ver_s16(short int *pixel_array, int nb_4_pixel_vectors, int row_byte_size, short int border[4]) # horizontal unsigned pixel conv (1/4 1/2 1/4) not tested # NOT TESTED
pixel_conv_ver_s16:
pushl %ebp movl %esp, %ebp push %esi push %edi
movl 8(%ebp), %esi # pixel array offset movl 12(%ebp), %ecx # nb of 4 pixel vectors in a row (at least 2) movl 16(%ebp), %edx # rowsize in bytes
movl 24(%ebp), %edi # mask vector movq (%edi), %mm5 movq 8(%edi), %mm6 movq 16(%edi), %mm7 movl 20(%ebp), %edi # edge vector
shll $1, %edx decl %ecx # loop has a terminator stub decl %ecx # loop has another terminator stub
movq (%edi), %mm0 # init regs (left edge, so mm0 is zero) movq (%esi), %mm1 movq (%esi,%edx,1), %mm2 jmp .conv_line_loop
.align 16 .conv_line_loop: call .conv_ver_4_pixels # compute conv movq %mm0, (%esi) # store result movq %mm1, %mm0 # mm0 <- prev (%esi) movq %mm2, %mm1 # mm1 <- (%esi,%edx,1) movq (%esi,%edx,2), %mm2 # mm2 <- (%esi,%edx,2) addl %edx, %esi # increase pointer decl %ecx jnz .conv_line_loop
call .conv_ver_4_pixels # compute conv movq %mm0, (%esi) # store result movq %mm1, %mm0 # mm0 <- prev (%esi) movq %mm2, %mm1 # mm1 <- (%esi,%edx,1) movq (%edi), %mm2 # clear invalid edge vector
addl %edx, %esi # increase pointer call .conv_ver_4_pixels # compute last vector movq %mm0, (%esi) # store it emms pop %edi pop %esi leave ret
--- NEW FILE: pixel_biquad_dirI_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. #
# TODO MOVE TO DIRECT FORM II # y[k] = b0 * x[k] + u1[k-1] # u1[k] = b1 * x[k] + u2[k-1] - a1 * y[k] # u2[k] = b2 * x[k] - a2 * y[k] # input in register: # %mm0-mm3: input 4x4 pixels {x0 x1 x2 x3} # %esi: coef memory (a1, a2, b0, b1, b2) # %edi: state memory (u1, u2)
# return in register: # %mm0-mm4: 4x4 pixels result
.biquad_4x4_pixels: .align 16 # prescale movq -8(%esi), %mm4 pmulhw %mm4, %mm0 pmulhw %mm4, %mm1 pmulhw %mm4, %mm2 pmulhw %mm4, %mm3 psllw $1, %mm0 psllw $1, %mm1 psllw $1, %mm2 psllw $1, %mm3
# first vector movq 0(%edi), %mm4 # mm4 <- u[-1] movq 8(%edi), %mm5 # mm5 <- u[-2] movq %mm4, %mm6 movq %mm5, %mm7
pmulhw 0(%esi), %mm6 # multiply by a1 pmulhw 8(%esi), %mm7 # multiply by a2
paddsw %mm6, %mm0 # accumulate paddsw %mm7, %mm0 # accumulate paddsw %mm0, %mm0 # scale by 2 (since all fixed point muls are x*y/2)
movq %mm0, %mm6 # mm6 <- u[0] movq %mm4, %mm7 # mm7 <- u[-1] pmulhw 16(%esi), %mm0 # multiply by b0 pmulhw 24(%esi), %mm4 # multiply by b1 pmulhw 32(%esi), %mm5 # multiply by b2
paddsw %mm4, %mm0 # accumulate paddsw %mm5, %mm0 # accumulate
# mm0 is result 0
# second vector movq %mm6, %mm4 # mm4 <- u[0] movq %mm7, %mm5 # mm5 <- u[-1]
pmulhw 0(%esi), %mm6 # multiply by a1 pmulhw 8(%esi), %mm7 # multiply by a2
paddsw %mm6, %mm1 # accumulate paddsw %mm7, %mm1 # accumulate paddsw %mm1, %mm1 # scale by 2
movq %mm1, %mm6 # mm6 <- u[1] movq %mm4, %mm7 # mm7 <- u[0] pmulhw 16(%esi), %mm1 # multiply by b0 pmulhw 24(%esi), %mm4 # multiply by b1 pmulhw 32(%esi), %mm5 # multiply by b2
paddsw %mm4, %mm1 # accumulate paddsw %mm5, %mm1 # accumulate
# mm1 is result 1
# third vector movq %mm6, %mm4 # mm4 <- u[1] movq %mm7, %mm5 # mm5 <- u[0]
pmulhw 0(%esi), %mm6 # multiply by a1 pmulhw 8(%esi), %mm7 # multiply by a2
paddsw %mm6, %mm2 # accumulate paddsw %mm7, %mm2 # accumulate paddsw %mm2, %mm2 # scale by 2
movq %mm2, %mm6 # mm6 <- u[2] movq %mm4, %mm7 # mm7 <- u[1] pmulhw 16(%esi), %mm2 # multiply by b0 pmulhw 24(%esi), %mm4 # multiply by b1 pmulhw 32(%esi), %mm5 # multiply by b2
paddsw %mm4, %mm2 # accumulate paddsw %mm5, %mm2 # accumulate
# mm2 is result 2
# fourth vector movq %mm6, %mm4 # mm4 <- u[2] movq %mm7, %mm5 # mm5 <- u[1]
pmulhw 0(%esi), %mm6 # multiply by a1 pmulhw 8(%esi), %mm7 # multiply by a2
paddsw %mm6, %mm3 # accumulate paddsw %mm7, %mm3 # accumulate paddsw %mm3, %mm3 # scale by 2
movq %mm3, 0(%edi) # store u[3] movq %mm4, 8(%edi) # store u[2] pmulhw 16(%esi), %mm3 # multiply by b0 pmulhw 24(%esi), %mm4 # multiply by b1 pmulhw 32(%esi), %mm5 # multiply by b2
paddsw %mm4, %mm3 # accumulate paddsw %mm5, %mm3 # accumulate
# mm3 is result 3
ret
# in order to use the 4 line parallel biquad routine on horizontal # lines, we need to reorder (rotate or transpose) the matrix, since # images are scanline encoded, and we want to work in parallell # on 4 lines. # # since the 4 lines are independent, it doesnt matter in which order # the the vector elements are present. # # this allows us to use the same routine for left->right and right->left # processing. # # some comments on the non-abelean group of square isometries consisting of # (I) identity # (H) horizontal axis mirror # (V) vertical axis mirror # (T) transpose (diagonal axis mirror) # (A) antitranspose (antidiagonal axis mirror) # (R1) 90deg anticlockwize rotation # (R2) 180deg rotation # (R3) 90deg clockwize rotation # # # we basicly have two options: (R1,R3) or (T,A) # we opt for T and A because they are self inverting, which improves locality # # use antitranspose for right to left an transpose # for left to right (little endian)
# antitranspose 4x4
# input # %mm3 == {d0 d1 d2 d3} # %mm2 == {c0 c1 c2 c3} # %mm1 == {b0 b1 b2 b3} # %mm0 == {a0 a1 a2 a3}
# output # %mm3 == {a3 b3 c3 d3} # %mm2 == {a2 b2 c2 d2} # %mm1 == {a1 b1 c1 d1} # %mm0 == {a0 b0 c0 d0}
.antitranspose_4x4: .align 16 movq %mm3, %mm4 punpcklwd %mm1, %mm4 # mm4 <- {b2 d2 b3 d3} movq %mm3, %mm5 punpckhwd %mm1, %mm5 # mm5 <- {b0 d0 b1 d1} movq %mm2, %mm6 punpcklwd %mm0, %mm6 # mm6 <- {a2 c2 a3 c3} movq %mm2, %mm7 punpckhwd %mm0, %mm7 # mm7 <- {a0 c0 a1 c1}
movq %mm4, %mm3 punpcklwd %mm6, %mm3 # mm3 <- {a3 b3 c3 d3} movq %mm4, %mm2 punpckhwd %mm6, %mm2 # mm2 <- {a2 b2 c2 d2} movq %mm5, %mm1 punpcklwd %mm7, %mm1 # mm1 <- {a1 b1 c1 d1} movq %mm5, %mm0 punpckhwd %mm7, %mm0 # mm0 <- {a0 b0 c0 d0}
ret
# transpose 4x4
# input # %mm3 == {d3 d2 d1 d0} # %mm2 == {c3 c2 c1 c0} # %mm1 == {b3 b2 b1 b0} # %mm0 == {a3 a2 a1 a0}
# output # %mm3 == {d3 c3 b3 a3} # %mm2 == {d2 c2 b2 a2} # %mm1 == {d1 c1 b1 a1} # %mm0 == {d0 c0 b0 a0}
.transpose_4x4: .align 16 movq %mm0, %mm4 punpcklwd %mm2, %mm4 # mm4 <- {c1 a1 c0 a0} movq %mm0, %mm5 punpckhwd %mm2, %mm5 # mm5 <- {c3 a3 c2 a2} movq %mm1, %mm6 punpcklwd %mm3, %mm6 # mm6 <- {d1 b1 d0 b0} movq %mm1, %mm7 punpckhwd %mm3, %mm7 # mm7 <- {d3 b3 d2 b2}
movq %mm4, %mm0 punpcklwd %mm6, %mm0 # mm0 <- {d0 c0 b0 a0} movq %mm4, %mm1 punpckhwd %mm6, %mm1 # mm1 <- {d1 c1 b1 a1} movq %mm5, %mm2 punpcklwd %mm7, %mm2 # mm2 <- {d2 c2 b2 a2} movq %mm5, %mm3 punpckhwd %mm7, %mm3 # mm3 <- {d3 c3 b3 a3}
ret
.globl pixel_biquad_vertb_s16 .type pixel_biquad_vertb_s16,@function
# pixel_biquad_vertbr_s16(char *pixel_array, int nb_rows, int linewidth, short int coef[20], short int state[8])
pixel_biquad_vertb_s16:
pushl %ebp movl %esp, %ebp push %ebx push %esi push %edi
movl 8(%ebp), %ebx # pixel array offset movl 12(%ebp), %ecx # nb of 4x4 pixblocks movl 16(%ebp), %edx # line with
movl 20(%ebp), %esi # coefs movl 24(%ebp), %edi # state
shll $1, %edx # short int addressing movl %edx, %eax shll $1, %eax addl %edx, %eax # eax = 3 * edx .align 16 .biquad_vertb_line_loop: movq (%ebx), %mm0 movq (%ebx,%edx,1), %mm1 movq (%ebx,%edx,2), %mm2 movq (%ebx,%eax,1), %mm3 call .biquad_4x4_pixels movq %mm0, (%ebx) movq %mm1, (%ebx,%edx,1) movq %mm2, (%ebx,%edx,2) movq %mm3, (%ebx,%eax,1) addl %edx, %ebx addl %eax, %ebx decl %ecx jnz .biquad_vertb_line_loop emms pop %edi pop %esi pop %ebx leave ret
.globl pixel_biquad_horlr_s16 .type pixel_biquad_horlr_s16,@function
# pixel_biquad_hor_s16(char *pixel_array, int nb_rows, int linewidth, short int coef[20], short int state[8])
pixel_biquad_horlr_s16:
pushl %ebp movl %esp, %ebp push %ebx push %esi push %edi
movl 8(%ebp), %ebx # pixel array offset movl 12(%ebp), %ecx # nb of 4x4 pixblocks movl 16(%ebp), %edx # line with
movl 20(%ebp), %esi # coefs movl 24(%ebp), %edi # state
shll $1, %edx # short int addressing movl %edx, %eax shll $1, %eax addl %edx, %eax # eax = 3 * edx .align 16 .biquad_horlr_line_loop: movq (%ebx), %mm0 movq (%ebx,%edx,1), %mm1 movq (%ebx,%edx,2), %mm2 movq (%ebx,%eax,1), %mm3 call .transpose_4x4 call .biquad_4x4_pixels call .transpose_4x4 movq %mm0, (%ebx) movq %mm1, (%ebx,%edx,1) movq %mm2, (%ebx,%edx,2) movq %mm3, (%ebx,%eax,1) addl $8, %ebx decl %ecx jnz .biquad_horlr_line_loop emms pop %edi pop %esi pop %ebx leave ret
--- NEW FILE: Makefile --- include ../../Makefile.config
OBJ = \ pixel_pack_s16u8.o \ pixel_unpack_u8s16.o \ pixel_add_s16.o \ pixel_mul_s16.o \ pixel_mix_s16.o \ pixel_randmix_s16.o \ pixel_conv_hor_s16.o \ pixel_conv_ver_s16.o \ pixel_biquad_s16.o \ pixel_ca_s1.o \ pixel_rand_s16.o \ pixel_crot_s16.o \ pixel_gain_s16.o \ pixel_resample_s16.o \ pixel_cheby_s16.o
all: $(OBJ)
test: pdp_mmx_test.o $(OBJ) gcc -o pdp_mmx_test pdp_mmx_test.o $(OBJ) -g
clean: rm -f *.o rm -f *~ rm -f pdp_mmx.a rm -f pdp_mmx_test
--- NEW FILE: pixel_add_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # .globl pixel_add_s16 .type pixel_add_s16,@function
# simple add # void pixel_add_s16(int *left, int *right, int nb_4pixel_vectors)
pixel_add_s16: pushl %ebp movl %esp, %ebp push %esi push %edi
movl 8(%ebp), %edi # left array movl 12(%ebp), %esi # right array movl 16(%ebp), %ecx # pixel count
.align 16 .loop_mix:
# prefetch 128(%esi) movq (%esi), %mm1 # load right 4 pixels from memory movq (%edi), %mm0 # load 4 left pixels from memory paddsw %mm1, %mm0 # mix movq %mm0, (%edi) addl $8, %esi addl $8, %edi decl %ecx jnz .loop_mix # loop
emms
pop %edi pop %esi leave ret
--- NEW FILE: pixel_randmix_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # .globl pixel_randmix_s16 .type pixel_randmix_s16,@function
# mmx rgba pixel gain # void pixel_randmix_s16(int *left, int *right, int nb_4pixel_vectors, short int random_seed[4], short int threshold[4])
pixel_randmix_s16: pushl %ebp movl %esp, %ebp push %esi push %edi
movl 20(%ebp), %edi # int16[4] array of random seeds movq (%edi), %mm6
movl 24(%ebp), %edi # int16[4] array of thresholds movq (%edi), %mm7 movl 8(%ebp), %edi # left array movl 12(%ebp), %esi # right array movl 16(%ebp), %ecx # pixel count
pcmpeqw %mm3, %mm3 psrlw $15, %mm3 # get bit mask 4 times 0x0001 .align 16 .loop_randmix:
# prefetch 128(%esi) movq (%esi), %mm1 # load right 4 pixels from memory movq (%edi), %mm0 # load 4 left pixels from memory
movq %mm6, %mm2 # get random vector pcmpgtw %mm7, %mm2 # compare random vector with threshold movq %mm2, %mm5 pand %mm0, %mm2 # get left array's components pandn %mm1, %mm5 # get right array's components por %mm2, %mm5 movq %mm5, (%edi) # store pixels
movq %mm6, %mm4 # get random vector psrlw $15, %mm4 # get first component movq %mm6, %mm5 psrlw $14, %mm5 # get second component pxor %mm5, %mm4 movq %mm6, %mm5 psrlw $12, %mm5 # get third component pxor %mm5, %mm4 movq %mm6, %mm5 psrlw $3, %mm5 # get forth component pxor %mm5, %mm4
psllw $1, %mm6 # shift left original random vector pand %mm3, %mm4 # isolate new bit por %mm4, %mm6 # combine into new random vector addl $8, %esi addl $8, %edi decl %ecx jnz .loop_randmix # loop
movl 20(%ebp), %edi # int16[4] array of random seeds movq %mm6, (%edi) # store random seeds
emms pop %edi pop %esi leave ret
--- NEW FILE: pixel_unpack_u8s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # .globl pixel_unpack_u8s16_y .type pixel_unpack_u8s16_y,@function
# mmx rgba pixel gain # void pixel_unpack_u8s16_y(char *input, char *output, int32 nb_pixels_div8)
pixel_unpack_u8s16_y: pushl %ebp movl %esp, %ebp push %esi push %edi
# movl 20(%ebp), %edi # int16[4] array of gains # movq (%edi), %mm7 # get gain array movl 8(%ebp), %esi # input uint8 pixel array movl 12(%ebp), %edi # output sint16 pixel array movl 16(%ebp), %ecx # nb of elements div 8
.align 16 .loop_unpack_y:
movq (%esi), %mm5 # load 8 pixels from memory pxor %mm0, %mm0 # zero mm0 - mm3 pxor %mm1, %mm1 punpcklbw %mm5, %mm0 # unpack 1st 4 pixels punpckhbw %mm5, %mm1 # unpack 2nd 4 pixles psrlw $0x1, %mm0 # shift right to clear sign bit 9.7 psrlw $0x1, %mm1 # pmulhw %mm7, %mm0 # apply gain # pmulhw %mm7, %mm1 # paddsw %mm0, %mm0 # correct factor 2 # paddsw %mm1, %mm1 movq %mm0, (%edi) # store movq %mm1, 8(%edi) addl $8, %esi # increment source pointer addl $16, %edi # increment dest pointer decl %ecx jnz .loop_unpack_y # loop
emms pop %edi pop %esi leave ret .globl pixel_unpack_u8s16_uv .type pixel_unpack_u8s16_uv,@function pixel_unpack_u8s16_uv: pushl %ebp movl %esp, %ebp push %esi push %edi
# movl 20(%ebp), %edi # int16[4] array of gains # movq (%edi), %mm7 # get gain array
movl 8(%ebp), %esi # input uint8 pixel array movl 12(%ebp), %edi # output sint16 pixel array movl 16(%ebp), %ecx # nb of elements div 8
pcmpeqw %mm6, %mm6 psllw $15, %mm6 .align 16 .loop_unpack_uv:
movq (%esi), %mm5 # load 8 pixels from memory pxor %mm0, %mm0 # zero mm0 - mm3 pxor %mm1, %mm1 punpcklbw %mm5, %mm0 # unpack 1st 4 pixels punpckhbw %mm5, %mm1 # unpack 2nd 4 pixles pxor %mm6, %mm0 # flip sign bit (Cr and Cb are ofset by 128) pxor %mm6, %mm1 # pmulhw %mm7, %mm0 # apply gain # pmulhw %mm7, %mm1 # paddsw %mm0, %mm0 # correct factor 2 # paddsw %mm1, %mm1 movq %mm0, (%edi) # store movq %mm1, 8(%edi) addl $8, %esi # increment source pointer addl $16, %edi # increment dest pointer decl %ecx jnz .loop_unpack_uv # loop
emms pop %edi pop %esi leave ret
--- NEW FILE: pixel_pack_s16u8.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # .globl pixel_pack_s16u8_y .type pixel_pack_s16u8_y,@function
# mmx rgba pixel gain # void pixel_pack_s16u8_y(int *input, int *output, int nb_8pixel_vectors)
pixel_pack_s16u8_y: pushl %ebp movl %esp, %ebp push %esi push %edi
# movl 20(%ebp), %edi # int16[4] array of gains # movq (%edi), %mm7 # get gain array # psllw $1, %mm7 # adjust for shifted sign bit movl 8(%ebp), %esi # input array movl 12(%ebp), %edi # output array movl 16(%ebp), %ecx # pixel count
pxor %mm6, %mm6 .align 16 .loop_pack_y:
# prefetch 128(%esi) movq (%esi), %mm0 # load 4 pixels from memory # pmulhw %mm7, %mm0 # apply gain movq 8(%esi), %mm1 # load 4 pixels from memory # pmulhw %mm7, %mm1 # apply gain
# movq %mm0, %mm2 # pcmpgtw %mm6, %mm2 # mm2 > 0 ? 0xffff : 0 # pand %mm2, %mm0
# movq %mm1, %mm3 # pcmpgtw %mm6, %mm3 # mm3 > 0 ? 0xffff : 0 # pand %mm3, %mm1
# psllw $1, %mm0 # shift out sign bit # psllw $1, %mm1 # shift out sign bit
psraw $7, %mm0 # shift to lsb psraw $7, %mm1 # shift to lsb packuswb %mm1, %mm0 # pack & saturate to 8bit vector movq %mm0, (%edi) # store result in memory
addl $16, %esi # increment source pointer addl $8, %edi # increment dest pointer decl %ecx jnz .loop_pack_y # loop
emms pop %edi pop %esi leave ret .globl pixel_pack_s16u8_uv .type pixel_pack_s16u8_uv,@function
pixel_pack_s16u8_uv: pushl %ebp movl %esp, %ebp push %esi push %edi
# movl 20(%ebp), %edi # int16[4] array of gains # movq (%edi), %mm7 # get gain array movl 8(%ebp), %esi # pixel array offset movl 12(%ebp), %edi # nb of elements movl 16(%ebp), %ecx # pixel count
pcmpeqw %mm6, %mm6 psllw $15, %mm6 movq %mm6, %mm5 psrlw $8, %mm5 por %mm5, %mm6 # mm6 <- 8 times 0x80 .align 16 .loop_pack_uv:
# prefetch 128(%esi) movq (%esi), %mm0 # load 4 pixels from memory # pmulhw %mm7, %mm0 # apply gain movq 8(%esi), %mm1 # load 4 pixels from memory # pmulhw %mm7, %mm1 # apply gain
psraw $8, %mm0 # shift to msb psraw $8, %mm1 packsswb %mm1, %mm0 # pack & saturate to 8bit vector pxor %mm6, %mm0 # flip sign bits movq %mm0, (%edi) # store result in memory
addl $16, %esi # increment source pointer addl $8, %edi # increment dest pointer decl %ecx jnz .loop_pack_uv # loop
emms pop %edi pop %esi leave ret
--- NEW FILE: pixel_mix_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # .globl pixel_mix_s16 .type pixel_mix_s16,@function
# mmx rgba pixel gain # void pixel_mix_s16(int *left, int *right, int nb_4pixel_vectors, # short int gain_left[4], short int gain_right[4])
pixel_mix_s16: pushl %ebp movl %esp, %ebp push %esi push %edi
movl 20(%ebp), %edi # int16[4] array of gains movq (%edi), %mm6 # get left gain array
movl 24(%ebp), %edi # int16[4] array of gains movq (%edi), %mm7 # get right gain array movl 8(%ebp), %edi # left array movl 12(%ebp), %esi # right array movl 16(%ebp), %ecx # pixel count
.align 16 .loop_mix:
# prefetch 128(%esi) movq (%esi), %mm1 # load right 4 pixels from memory pmulhw %mm7, %mm1 # apply right gain movq (%edi), %mm0 # load 4 left pixels from memory pmulhw %mm6, %mm0 # apply left gain # pslaw $1, %mm1 # shift left ((s).15 x (s).15 -> (s0).14)) # pslaw $1, %mm0 paddsw %mm0, %mm0 # no shift left arithmic, so use add instead paddsw %mm1, %mm1 paddsw %mm1, %mm0 # mix movq %mm0, (%edi) addl $8, %esi addl $8, %edi decl %ecx jnz .loop_mix # loop
emms
pop %edi pop %esi leave ret
--- NEW FILE: pixel_ca_s1.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. #
# this file contains assembler routines for 2D 1 bit cellular automata # processing. it is organized around a feeder kernel and a # stack based bit processor (virtual forth machine) # # the feeder kernel is responsable for loading/storing CA cells # from/to memory. data in memory is organized as a scanline # encoded toroidial bitplane (lsb = left). to simplify the kernel, the top # left corner of the rectangular grid of pixels will shift down # every processing step. # # the stack machine has the following architecture: # CA stack: %esi, TOS: %mm0 (32x2 pixels. lsw = top row) # CA horizon: %mm4-%mm7 (64x4 pixels. %mm4 = top row) # # the stack size / organization is not known to the stack machine. # it can be thought of as operating on a 3x3 cell neightbourhood. # the only purpose of forth program is to determine the CA local update rule. # # the machine is supposed to be very minimal. no looping control. # no adressing modes. no conditional code (hey, this is an experiment!) # so recursion is not allowed (no way to stop it) # there are 9 words to load the cell neigbourhood on the stack. # the rest is just logic and stack manips.
# this file contains pure asm macros. it is to be included before assembly # after scaforth.pl has processed the .scaf file
# *************************** CA CELL ACCESS MACROS ***************************** # fetchTL - fetchBR
# shift / load rectangle macros:
# shift rectangle horizontal # result is in reg1 .macro shift reg1 reg2 count psllq $(32+\count), \reg1 psrlq $(32-\count), \reg2 psrlq $32, \reg1 psllq $32, \reg2 por \reg2, \reg1 .endm
.macro ldtop reg1 reg2 movq %mm4, \reg1 movq %mm5, \reg2 .endm
.macro ldcenter reg1 reg2 movq %mm5, \reg1 movq %mm6, \reg2 .endm
.macro ldbottom reg1 reg2 movq %mm6, \reg1 movq %mm7, \reg2 .endm
# fetch from top row
# fetch the top left square .macro fetchTL ldtop %mm0, %mm1 shift %mm0, %mm1, -1 .endm
# fetch the top mid square .macro fetchTM ldtop %mm0, %mm1 shift %mm0, %mm1, 0 .endm
# fetch the top right square .macro fetchTR ldtop %mm0, %mm1 shift %mm0, %mm1, 1 .endm
# fetch from center row
# fetch the mid left square .macro fetchML ldcenter %mm0, %mm1 shift %mm0, %mm1, -1 .endm
# fetch the mid mid square .macro fetchMM ldcenter %mm0, %mm1 shift %mm0, %mm1, 0 .endm
# fetch the mid right square .macro fetchMR ldcenter %mm0, %mm1 shift %mm0, %mm1, 1 .endm
# fetch from bottom row
# fetch the bottom left square .macro fetchBL ldbottom %mm0, %mm1 shift %mm0, %mm1, -1 .endm
# fetch the bottom mid square .macro fetchBM ldbottom %mm0, %mm1 shift %mm0, %mm1, 0 .endm
# fetch the bottom right square .macro fetchBR ldbottom %mm0, %mm1 shift %mm0, %mm1, 1 .endm
# *************************** CA STACK MANIP MACROS ***************************** # dup drop dropdup swap nip dropover
.macro dup lea -8(%esi), %esi movq %mm0, (%esi) .endm
.macro drop movq (%esi), %mm0 lea 8(%esi), %esi .endm
.macro dropdup movq (%esi), %mm0 .endm
.macro swap movq (%esi), %mm1 movq %mm0, (%esi) movq %mm1, %mm0 .endm
.macro nip lea 8(%esi), %esi .endm
.macro dropover movq 8(%esi), %mm0 .endm
# *************************** CA BOOLEAN LOGIC MACROS ***************************** # overxor .macro overxor pxor (%esi), %mm0 .endm
--- NEW FILE: pixel_biquad_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. #
# DIRECT FORM II BIQUAD # # y[k] = b0 * x[k] + u1[k-1] # u1[k] = b1 * x[k] + u2[k-1] - a1 * y[k] # u2[k] = b2 * x[k] - a2 * y[k] # MACRO: df2 <reg> # # computes a direct form 2 biquad # does not use {mm0-mm3}<inreg> # # input: <reg> == input # %mm4 == state 1 # %mm5 == state 2 # (%esi) == biquad coefs (-a1 -a2 b0 b1 b2) in s1.14 # output: <reg> == output # %mm4 == state 1 # %mm5 == state 2
.macro df2 reg movq \reg, %mm6 # mm6 == x[k] movq \reg, %mm7 # mm7 == x[k] pmulhw 16(%esi), %mm6 # mm6 == x[k] * b0 pmulhw 24(%esi), %mm7 # mm7 == x[k] * b1 paddw %mm4, %mm6 # mm6 == x[k] * b0 + u1[k-1] == y[k] paddw %mm5, %mm7 # mm7 == x[k] * b1 + u2[k-1] paddsw %mm6, %mm6 # compensate for mul = x*y/4 (coefs are s1.14 fixed point) paddsw %mm6, %mm6 # paddsw ensures saturation movq \reg, %mm5 # mm5 == x[k] movq %mm6, %mm4 # mm4 == y[k] movq %mm6, \reg # reg == y[k] -------------------- pmulhw 0(%esi), %mm4 # mm4 == y[k] * (-a1) pmulhw 8(%esi), %mm6 # mm6 == y[k] * (-a2) pmulhw 32(%esi), %mm5 # mm5 == x[k] * b2 paddw %mm7, %mm4 # mm4 == u1[k] -------------------- paddw %mm6, %mm5 # mm5 == u2[k] -------------------- .endm
# input in register: # %mm0-mm3: input 4x4 pixels {x0 x1 x2 x3} # %esi: coef memory (-a1, -a2, b0, b1, b2) in s1.14 # %edi: state memory (u1, u2) # return in register: # %mm0-mm4: 4x4 pixels result
.macro biquad_4x4_pixels .align 16 movq 0(%edi), %mm4 # get state movq 8(%edi), %mm5 df2 %mm0 # compute 4 biquads df2 %mm1 df2 %mm2 df2 %mm3 movq %mm4, 0(%edi) # store state movq %mm5, 8(%edi) .endm
# in order to use the 4 line parallel biquad routine on horizontal # lines, we need to reorder (rotate or transpose) the matrix, since # images are scanline encoded, and we want to work in parallell # on 4 lines. # # since the 4 lines are independent, it doesnt matter in which order # the the vector elements are present. # # this allows us to use the same routine for left->right and right->left # processing. # # some comments on the non-abelean group of square isometries consisting of # (I) identity # (H) horizontal axis mirror # (V) vertical axis mirror # (T) transpose (diagonal axis mirror) # (A) antitranspose (antidiagonal axis mirror) # (R1) 90deg anticlockwize rotation # (R2) 180deg rotation # (R3) 90deg clockwize rotation # # # we basicly have two options: (R1,R3) or (T,A) # we opt for T and A because they are self inverting, which improves locality # # use antitranspose for right to left an transpose # for left to right (little endian)
# antitranspose 4x4
# input # %mm3 == {d0 d1 d2 d3} # %mm2 == {c0 c1 c2 c3} # %mm1 == {b0 b1 b2 b3} # %mm0 == {a0 a1 a2 a3}
# output # %mm3 == {a3 b3 c3 d3} # %mm2 == {a2 b2 c2 d2} # %mm1 == {a1 b1 c1 d1} # %mm0 == {a0 b0 c0 d0}
.macro antitranspose_4x4: movq %mm3, %mm4 punpcklwd %mm1, %mm4 # mm4 <- {b2 d2 b3 d3} movq %mm3, %mm5 punpckhwd %mm1, %mm5 # mm5 <- {b0 d0 b1 d1} movq %mm2, %mm6 punpcklwd %mm0, %mm6 # mm6 <- {a2 c2 a3 c3} movq %mm2, %mm7 punpckhwd %mm0, %mm7 # mm7 <- {a0 c0 a1 c1}
movq %mm4, %mm3 punpcklwd %mm6, %mm3 # mm3 <- {a3 b3 c3 d3} movq %mm4, %mm2 punpckhwd %mm6, %mm2 # mm2 <- {a2 b2 c2 d2} movq %mm5, %mm1 punpcklwd %mm7, %mm1 # mm1 <- {a1 b1 c1 d1} movq %mm5, %mm0 punpckhwd %mm7, %mm0 # mm0 <- {a0 b0 c0 d0} .endm
# transpose 4x4
# input # %mm3 == {d3 d2 d1 d0} # %mm2 == {c3 c2 c1 c0} # %mm1 == {b3 b2 b1 b0} # %mm0 == {a3 a2 a1 a0}
# output # %mm3 == {d3 c3 b3 a3} # %mm2 == {d2 c2 b2 a2} # %mm1 == {d1 c1 b1 a1} # %mm0 == {d0 c0 b0 a0}
.macro transpose_4x4: movq %mm0, %mm4 punpcklwd %mm2, %mm4 # mm4 <- {c1 a1 c0 a0} movq %mm0, %mm5 punpckhwd %mm2, %mm5 # mm5 <- {c3 a3 c2 a2} movq %mm1, %mm6 punpcklwd %mm3, %mm6 # mm6 <- {d1 b1 d0 b0} movq %mm1, %mm7 punpckhwd %mm3, %mm7 # mm7 <- {d3 b3 d2 b2}
movq %mm4, %mm0 punpcklwd %mm6, %mm0 # mm0 <- {d0 c0 b0 a0} movq %mm4, %mm1 punpckhwd %mm6, %mm1 # mm1 <- {d1 c1 b1 a1} movq %mm5, %mm2 punpcklwd %mm7, %mm2 # mm2 <- {d2 c2 b2 a2} movq %mm5, %mm3 punpckhwd %mm7, %mm3 # mm3 <- {d3 c3 b3 a3}
.endm .globl pixel_biquad_vertb_s16 .type pixel_biquad_vertb_s16,@function
# pixel_biquad_vertbr_s16(char *pixel_array, int nb_rows, int linewidth, short int coef[20], short int state[8])
pixel_biquad_vertb_s16:
pushl %ebp movl %esp, %ebp push %ebx push %esi push %edi
movl 8(%ebp), %ebx # pixel array offset movl 12(%ebp), %ecx # nb of 4x4 pixblocks movl 16(%ebp), %edx # line with
movl 20(%ebp), %esi # coefs movl 24(%ebp), %edi # state
shll $1, %edx # short int addressing movl %edx, %eax shll $1, %eax addl %edx, %eax # eax = 3 * edx .align 16 .biquad_vertb_line_loop: movq (%ebx), %mm0 movq (%ebx,%edx,1), %mm1 movq (%ebx,%edx,2), %mm2 movq (%ebx,%eax,1), %mm3 biquad_4x4_pixels movq %mm0, (%ebx) movq %mm1, (%ebx,%edx,1) movq %mm2, (%ebx,%edx,2) movq %mm3, (%ebx,%eax,1) addl %edx, %ebx addl %eax, %ebx decl %ecx jnz .biquad_vertb_line_loop emms pop %edi pop %esi pop %ebx leave ret .globl pixel_biquad_verbt_s16 .type pixel_biquad_verbt_s16,@function
# pixel_biquad_vertbt_s16(char *pixel_array, int nb_rows, int linewidth, short int coef[20], short int state[8])
pixel_biquad_verbt_s16:
pushl %ebp movl %esp, %ebp push %ebx push %esi push %edi
movl 8(%ebp), %ebx # pixel array offset movl 12(%ebp), %ecx # nb of 4x4 pixblocks movl 16(%ebp), %eax # line with
shll $3, %eax # 4 line byte spacing decl %ecx mul %ecx incl %ecx addl %eax, %ebx # ebx points to last pixblock
movl 16(%ebp), %edx # line with
movl 20(%ebp), %esi # coefs movl 24(%ebp), %edi # state
shll $1, %edx # short int addressing movl %edx, %eax shll $1, %eax addl %edx, %eax # eax = 3 * edx .align 16 .biquad_verbt_line_loop: movq (%ebx), %mm3 movq (%ebx,%edx,1), %mm2 movq (%ebx,%edx,2), %mm1 movq (%ebx,%eax,1), %mm0 biquad_4x4_pixels movq %mm3, (%ebx) movq %mm2, (%ebx,%edx,1) movq %mm1, (%ebx,%edx,2) movq %mm0, (%ebx,%eax,1) subl %edx, %ebx subl %eax, %ebx decl %ecx jnz .biquad_verbt_line_loop emms pop %edi pop %esi pop %ebx leave ret
.globl pixel_biquad_horlr_s16 .type pixel_biquad_horlr_s16,@function # pixel_biquad_hor_s16(char *pixel_array, int nb_rows, int linewidth, short int coef[20], short int state[8])
pixel_biquad_horlr_s16:
pushl %ebp movl %esp, %ebp push %ebx push %esi push %edi
movl 8(%ebp), %ebx # pixel array offset movl 12(%ebp), %ecx # nb of 4x4 pixblocks movl 16(%ebp), %edx # line with
movl 20(%ebp), %esi # coefs movl 24(%ebp), %edi # state
shll $1, %edx # short int addressing movl %edx, %eax shll $1, %eax addl %edx, %eax # eax = 3 * edx .align 16 .biquad_horlr_line_loop: movq (%ebx), %mm0 movq (%ebx,%edx,1), %mm1 movq (%ebx,%edx,2), %mm2 movq (%ebx,%eax,1), %mm3 transpose_4x4 biquad_4x4_pixels transpose_4x4 movq %mm0, (%ebx) movq %mm1, (%ebx,%edx,1) movq %mm2, (%ebx,%edx,2) movq %mm3, (%ebx,%eax,1) addl $8, %ebx decl %ecx jnz .biquad_horlr_line_loop emms pop %edi pop %esi pop %ebx leave ret
.globl pixel_biquad_horrl_s16 .type pixel_biquad_horrl_s16,@function # pixel_biquad_horrl_s16(char *pixel_array, int nb_rows, int linewidth, short int coef[20], short int state[8])
pixel_biquad_horrl_s16:
pushl %ebp movl %esp, %ebp push %ebx push %esi push %edi
movl 8(%ebp), %ebx # pixel array offset movl 12(%ebp), %ecx # nb of 4x4 pixblocks movl 16(%ebp), %edx # line with
movl %ecx, %eax decl %eax shll $3, %eax addl %eax, %ebx # ebx points to last pixblock
movl 20(%ebp), %esi # coefs movl 24(%ebp), %edi # state
shll $1, %edx # short int addressing movl %edx, %eax shll $1, %eax addl %edx, %eax # eax = 3 * edx .align 16 .biquad_horrl_line_loop: movq (%ebx), %mm0 movq (%ebx,%edx,1), %mm1 movq (%ebx,%edx,2), %mm2 movq (%ebx,%eax,1), %mm3 antitranspose_4x4 biquad_4x4_pixels antitranspose_4x4 movq %mm0, (%ebx) movq %mm1, (%ebx,%edx,1) movq %mm2, (%ebx,%edx,2) movq %mm3, (%ebx,%eax,1) subl $8, %ebx decl %ecx jnz .biquad_horrl_line_loop emms pop %edi pop %esi pop %ebx leave ret
.globl pixel_biquad_time_s16 .type pixel_biquad_time_s16,@function # pixel_biquad_time_s16(short int *pixel_array, short int *s1, short int *s2, short int *coefs, int nb_4_pix_vectors)
pixel_biquad_time_s16:
pushl %ebp movl %esp, %ebp push %ebx push %esi push %edi
movl 8(%ebp), %ebx # pixel array offset movl 12(%ebp), %edx # state 1 array movl 16(%ebp), %edi # state 2 array
movl 20(%ebp), %esi # coefs movl 24(%ebp), %ecx # nb of 4 pixel vectors
.align 16 .biquad_time_loop: movq (%ebx), %mm0 # get input movq (%edx), %mm4 # get state 1 movq (%edi), %mm5 # get state 2 df2 %mm0 # compute direct form 2 movq %mm0, (%ebx) # write output movq %mm5, (%edi) # write state 2 movq %mm4, (%edx) # write state 1 addl $8, %ebx addl $8, %edi addl $8, %edx decl %ecx jnz .biquad_time_loop emms pop %edi pop %esi pop %ebx leave ret
--- NEW FILE: pixel_conv_hor_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # # intermediate function # input in register: # %mm0: left 4 pixels # %mm1: middle 4 pixels # %mm2: right 4 pixels # %mm5: left 4 pixel masks # %mm6: middle 4 pixel masks # %mm7: right 4 pixel masks # return in register: # %mm0: middle 4 pixels result
.conv_hor_4_pixels: .align 16 # compute quadruplet
# get left pixels psrlq $48, %mm0 # shift word 3 to byte 0 movq %mm1, %mm4 psllq $16, %mm4 # shift word 0,1,2 to 1,2,3 por %mm4, %mm0 # combine pmulhw %mm5, %mm0 psllw $1, %mm0
# get middle pixels movq %mm1, %mm4 pmulhw %mm6, %mm4 psllw $1, %mm4 paddsw %mm4, %mm0
# get right pixels movq %mm2, %mm3 psllq $48, %mm3 # shift word 0 to word 3 movq %mm1, %mm4 psrlq $16, %mm4 # shift word 1,2,3 to 0,1,2 por %mm4, %mm3 # combine pmulhw %mm7, %mm3 psllw $1, %mm3 paddsw %mm3, %mm0 # accumulate ret .globl pixel_conv_hor_s16 .type pixel_conv_hor_s16,@function
# pixel_conv_hor_s16(short int *pixel_array, int nb_4_pixel_vectors, short int border[4], short int mask[12]) # horizontal unsigned pixel conv (1/4 1/2 1/4) not tested # NOT TESTED
pixel_conv_hor_s16:
pushl %ebp movl %esp, %ebp push %esi push %edi
movl 8(%ebp), %esi # pixel array offset movl 12(%ebp), %ecx # nb of 8 pixel vectors in a row (at least 2)
movl 20(%ebp), %edi # mask vector movq (%edi), %mm5 movq 8(%edi), %mm6 movq 16(%edi), %mm7 movl 16(%ebp), %edi # boundary pixel vector
movq (%edi), %mm0 # init regs (left edge, so mm0 is zero) movq (%esi), %mm1 movq 8(%esi), %mm2
decl %ecx # loop has 2 terminator stubs decl %ecx # todo: handle if ecx < 3 jmp .conv_line_loop
.align 16 .conv_line_loop: call .conv_hor_4_pixels # compute conv movq %mm0, (%esi) # store result movq %mm1, %mm0 # mm0 <- prev (%esi) movq %mm2, %mm1 # mm1 <- 8(%esi) movq 16(%esi), %mm2 # mm2 <- 16(%esi) addl $8, %esi # increase pointer decl %ecx jnz .conv_line_loop
call .conv_hor_4_pixels # compute conv movq %mm0, (%esi) # store result movq %mm1, %mm0 # mm0 <- prev (%esi) movq %mm2, %mm1 # mm1 <- 8(%esi) movq (%edi), %mm2 # mm2 <- border
call .conv_hor_4_pixels # compute last vector movq %mm0, 8(%esi) # store it emms pop %edi pop %esi leave ret
--- NEW FILE: pixel_crot_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # .globl pixel_crot3d_s16 .type pixel_crot3d_s16,@function
# 3 dimensional colour space rotation # 3x3 matrix is column encoded, each coefficient is a 4x16 bit fixed point vector # void pixel_crot3d_s16(int *buf, int nb_4pixel_vectors_per_plane, short int *matrix)
pixel_crot3d_s16: pushl %ebp movl %esp, %ebp push %esi push %edi
movl 8(%ebp), %esi # input array movl 12(%ebp), %ecx # pixel count movl 16(%ebp), %edi # rotation matrix movl %ecx, %edx shll $3, %edx # %edx = plane spacing
.align 16 .loop_crot3d:
movq (%esi), %mm0 # get 1st component movq (%esi,%edx,1), %mm6 # get 2nd component movq (%esi,%edx,2), %mm7 # get 3rd component
movq %mm0, %mm1 # copy 1st component movq %mm0, %mm2
pmulhw (%edi), %mm0 # mul first column pmulhw 8(%edi), %mm1 pmulhw 16(%edi), %mm2
movq %mm6, %mm5 # copy 2nd component movq %mm6, %mm3
pmulhw 24(%edi), %mm6 # mul second column pmulhw 32(%edi), %mm5 pmulhw 40(%edi), %mm3
paddsw %mm6, %mm0 # accumulate paddsw %mm5, %mm1 paddsw %mm3, %mm2
movq %mm7, %mm4 # copy 3rd component movq %mm7, %mm6
pmulhw 48(%edi), %mm4 # mul third column pmulhw 56(%edi), %mm6 pmulhw 64(%edi), %mm7
paddsw %mm4, %mm0 # accumulate paddsw %mm6, %mm1 paddsw %mm7, %mm2
paddsw %mm0, %mm0 # double (fixed point normalization) paddsw %mm1, %mm1 paddsw %mm2, %mm2
movq %mm0, (%esi) # store movq %mm1, (%esi, %edx, 1) movq %mm2, (%esi, %edx, 2)
addl $8, %esi # increment source pointer decl %ecx jnz .loop_crot3d # loop
emms pop %edi pop %esi leave ret
.globl pixel_crot2d_s16 .type pixel_crot2d_s16,@function # 2 dimensional colour space rotation # 2x2 matrix is column encoded, each coefficient is a 4x16 bit fixed point vector # void pixel_crot2d_s16(int *buf, int nb_4pixel_vectors_per_plane, short int *matrix)
pixel_crot2d_s16: pushl %ebp movl %esp, %ebp push %esi push %edi
movl 8(%ebp), %esi # input array movl 12(%ebp), %ecx # pixel count movl 16(%ebp), %edi # rotation matrix movl %ecx, %edx shll $3, %edx # %edx = plane spacing
.align 16 .loop_crot2d:
movq (%esi), %mm0 # get 1st component movq (%esi,%edx,1), %mm2 # get 2nd component
movq %mm0, %mm1 # copy 1st component movq %mm2, %mm3 # copy 2nd component
pmulhw (%edi), %mm0 # mul first column pmulhw 8(%edi), %mm1
pmulhw 16(%edi), %mm2 # mul second column pmulhw 24(%edi), %mm3
paddsw %mm2, %mm0 # accumulate paddsw %mm3, %mm1
paddsw %mm0, %mm0 # fixed point gain correction paddsw %mm1, %mm1
movq %mm0, (%esi) # store movq %mm1, (%esi, %edx, 1)
addl $8, %esi # increment source pointer decl %ecx jnz .loop_crot2d # loop
emms pop %edi pop %esi leave ret
--- NEW FILE: pixel_s1.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. #
# this file contains ops for binary image processing # 8x8 bit tile encoded # low byte = bottom row # low bit = right column # %mm7 = scratch reg for all macros
# ************ load mask ******************* # compute bit masks for rows and columns # %mm7: scratch reg
# load mask top .macro ldmt count reg pcmpeqb \reg, \reg psllq $(64-(\count<<3)), \reg .endm
# load mask bottom .macro ldmb count reg pcmpeqb \reg, \reg psrlq $(64-(\count<<3)), \reg .endm
# load mask top and bottom .macro ldmtb count regt regb ldmb \count, \regb ldmt \count, \regt .endm
# load mask right .macro ldmr count reg pcmpeqb %mm7, %mm7 psrlw $(16-\count), %mm7 movq %mm7, \reg psllq $8, %mm7 por %mm7, \reg .endm
# load mask left .macro ldml count reg pcmpeqb %mm7, %mm7 psllw $(16-\count), %mm7 movq %mm7, \reg psrlq $8, %mm7 por %mm7, \reg .endm
# load mask left and right .macro ldmlr count regl regr pcmpeqb %mm7, %mm7 psllw $(16-\count), %mm7 movq %mm7, \regl psrlq $8, %mm7 por %mm7, \regl movq \regl, \regr psrlq $(8-\count), \regr .endm
# ************* shift square ********** # shifts a square in reg, fills with zeros
# shift square top .macro sst count reg psllq $(\count<<3), \reg .endm
# shift square bottom .macro ssb count reg psrlq $(\count<<3), \reg .endm
# not tested # shift square left .macro ssl count reg movq \reg, %mm7 pcmpeqb \reg, \reg psllw $(16-\count), \reg psrlw $8, \reg pandn %mm7, \reg psllw $(\count), \reg .endm
# shift square right .macro ssr count reg movq \reg, %mm7 pcmpeqb \reg, \reg psrlw $(16-\count), \reg psllw $8, \reg pandn %mm7, \reg psrlw $(\count), \reg .endm
# ********** combine square ************* # combines 2 squares
# combine right .macro csr count regr reg ssl \count, \reg ssr (8-\count), \regr por \regr, \reg .endm
# combine left .macro csl count regl reg ssr \count, \reg ssl (8-\count), \regl por \regl, \reg .endm
# combine top .macro cst count regt reg ssb \count, \reg sst (8-\count), \regt por \regt, \reg .endm
# combine bottom .macro csb count regb reg sst \count, \reg ssb (8-\count), \regb por \regb, \reg .endm
# ********** load combine square ************* # loads combined square using mask
# load combined square left # mask should be count bits set right (i.e. 0x01) .macro lcsml count mask source sourcel dstreg movq \mask, \dstreg movq \mask, %mm7 pandn \source, \dstreg pand \sourcel, %mm7 psrlq $(\count), \dstreg psllq $(8-\count), %mm7 por %mm7, \dstreg .endm .globl pixel_test_s1 .type pixel_test_s1,@function
# simple add # void pixel_add_s16(void *dest, void *source, int nb_squares, int spacing)
#
pixel_test_s1: pushl %ebp movl %esp, %ebp push %esi push %edi
movl 8(%ebp), %edi # dest movl 12(%ebp), %esi # source movl 16(%ebp), %ecx # count movl 20(%ebp), %edx # row distance
ldmr 1, %mm6 lcsml 1, %mm6, (%esi), 8(%esi), %mm0 movq %mm0, (%edi)
# movq (%esi), %mm0 # movq 8(%esi), %mm1 # csl 4, %mm1, %mm0 # movq %mm0, (%edi)
emms
pop %edi pop %esi leave ret
--- NEW FILE: pdp_mmx_test.c --- #include "pdp_mmx.h"
#define FP(x) ((short int)(((float)(x) * 2 * 256.0f)))
#define nbp 256
short int a1[4] = {0x0100,0x0100,0x0100,0x0100}; short int a2[4] = {0x0100,0x0100,0x0100,0x0100}; short int b0[4] = {0x0100,0x0100,0x0100,0x0100}; short int b1[4] = {0x0100,0x0100,0x0100,0x0100}; short int b2[4] = {0x0100,0x0100,0x0100,0x0100};
short int u1[4] = {0x0100,0x0100,0x0100,0x0100}; short int u2[4] = {0x0100,0x0100,0x0100,0x0100};
short int x0[4] = {0x0100,0x0100,0x0100,0x0100}; short int x1[4] = {0x0100,0x0100,0x0100,0x0100}; short int x2[4] = {0x0100,0x0100,0x0100,0x0100}; short int x3[4] = {0x0100,0x0100,0x0100,0x0100};
void print_pixel(unsigned int i) { if (i) printf("x "); else printf(". "); }
void print_line(void) { printf("\n"); }
void print_square(unsigned char *c) { int i,j;
for(j=7; j>=0; j--){ for(i=0; i<8; i++) print_pixel(c[j] & (1<<(7-i))); printf("\n"); }
}
main() {
unsigned char src[16]={1,2,3,4,5,6,7,8,-1,-2,-3,-4,-5,-6,-7,-8}; unsigned char dst[8];
print_square(src); print_line(); print_square(src+8); print_line();
pixel_test_s1(dst,src,1,1);
print_square(dst); print_line();
}
--- NEW FILE: pixel_mul_s16.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # .globl pixel_mul_s16 .type pixel_mul_s16,@function
# simple add # void pixel_mul_s16(int *left, int *right, int nb_4pixel_vectors)
pixel_mul_s16: pushl %ebp movl %esp, %ebp push %esi push %edi
movl 8(%ebp), %edi # left array movl 12(%ebp), %esi # right array movl 16(%ebp), %ecx # pixel count
.align 16 .loop_mix:
# prefetch 128(%esi) movq (%esi), %mm1 # load right 4 pixels from memory movq (%edi), %mm0 # load 4 left pixels from memory pmulhw %mm1, %mm0 # mul psllw $1, %mm0 # fixed point shift correction movq %mm0, (%edi) addl $8, %esi addl $8, %edi decl %ecx jnz .loop_mix # loop
emms
pop %edi pop %esi leave ret
--- NEW FILE: pixel_gain.s --- # Pure Data Packet mmx routine. # Copyright (c) by Tom Schouten pdp@zzz.kotnet.org # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. # .globl pixel_gain .type pixel_gain,@function
# mmx rgba pixel gain # void asmtest(char *pixelarray, int32 nbpixels, int *rgba_gain) # gains are 7.9 fixed point for rgba
pixel_gain: pushl %ebp movl %esp, %ebp push %esi push %edi
movl 8(%ebp), %esi # pixel array offset movl 12(%ebp), %ecx # nb of elements movl 16(%ebp), %edi # int16[4] array of gains
prefetch (%esi)
emms sarl $2, %ecx # process 4 pixels per loop iteration jz .exit movq (%edi), %mm7 # read gain array from memory jmp .loop_gain
.align 16 .loop_gain:
prefetch 128(%esi) movq (%esi), %mm5 # load pixel 1-2 from memory movq 8(%esi), %mm6 # load pixel 3-4 from memory pxor %mm0, %mm0 # zero mm0 - mm3 pxor %mm1, %mm1 pxor %mm2, %mm2 pxor %mm3, %mm3 punpcklbw %mm5, %mm0 # unpack 1st pixel into 8.8 bit ints punpckhbw %mm5, %mm1 # unpack 2nd punpcklbw %mm6, %mm2 # unpack 3rd punpckhbw %mm6, %mm3 # unpack 4th psrlw $0x1, %mm0 # shift right to clear sign bit 9.7 psrlw $0x1, %mm1 psrlw $0x1, %mm2 psrlw $0x1, %mm3 pmulhw %mm7, %mm0 # multiply 1st pixel 9.7 * 7.9 -> 16.0 pmulhw %mm7, %mm1 # multiply 2nd pmulhw %mm7, %mm2 # multiply 3rd pmulhw %mm7, %mm3 # multiply 4th
packuswb %mm1, %mm0 # pack & saturate to 8bit vector movq %mm0, (%esi) # store result in memory packuswb %mm3, %mm2 # pack & saturate to 8bit vector movq %mm2, 8(%esi) # store result in memory
addl $16, %esi # increment source pointer decl %ecx jnz .loop_gain # loop
.exit: emms pop %edi pop %esi leave ret